Teaching robots to feel their way through complex tasks

Tactile sensing—touch feedback—is underused in robot learning despite being how humans manipulate objects without looking. This benchmark standardizes tactile reinforcement learning across four robot designs and focuses on truly blind manipulation using only touch and joint position. Agents tuned on the platform perform dexterous tasks like spinning Baoding balls 13× faster than prior methods, and the team released code and baseline models to let researchers focus on algorithms rather than tedious hyperparameter tuning.